I am a Computer Science undergraduate (Data Science minor) at Drew University, working at the intersection of Large Language Models, Retrieval-Augmented Generation, and Computer Vision.
My focus is narrow on purpose: I build domain-specific AI systems that actually ship — crawled corpora, fine-tuned models, retrieval pipelines, and deployed apps that a non-technical person can open in a browser and use.
| 🔬 Currently | Undergraduate Researcher, Drew Summer Science Institute — Fermilab physics Q&A with RAG + QLoRA |
| 🎯 Core focus | LLM fine-tuning (QLoRA/PEFT) · Hybrid retrieval (BM25 + dense) · Computer Vision · MLOps |
| 🛠️ I like building | AI agents, retrieval systems, automation tools, security-testing utilities |
| 🌎 Based in | Madison, New Jersey, USA |
| 🎓 Graduating | May 2027 · Open to Data Science / ML Engineering internships |
| 📚 Always | learning, improving, and experimenting |
| Achievement | Detail |
|---|---|
| 🥇 DSSI Research Fellow, 2026 | Selected for a competitive 8-week funded summer research fellowship (full scholarship + 4 academic credits) |
| 🎤 Day of Scholars, Spring 2027 | Selected to present research at Drew University's annual scholarship showcase (scheduled) |
| 🤖 936 pages → 8,847 chunks → fine-tuned LLM | Built a full domain-specific RAG stack end-to-end, from crawler to deployed UI |
| 📈 +19 pts token accuracy | Raised fine-tuned model token-level accuracy from 56% → ~75% with no overfitting |
| 🎓 GPA 3.75 / 4.0 | Merit Scholarship $15,000 (Fall 2025 – Spring 2026) |
| 👨🏫 CRLA Level I Certified Tutor | Peer Subject Tutor in CS & Data Science at Drew's Center for Academic Excellence |
- Selected for a competitive, 8-week funded summer research fellowship under direct faculty mentorship.
- Award package: full scholarship + 4 academic credits.
- Placed on a live LLM research project (Fermilab physics Q&A) rather than a coursework-style assignment.
- Selected to present DSSI research at Drew's annual university-wide scholarship showcase (scheduled).
- Selection is based on the strength and completeness of the summer research output.
- Awarded for academic performance, Fall 2025 – Spring 2026.
- GPA 3.75 / 4.0 in B.S. Computer Science (Data Science minor).
- Selected as a Peer Subject Tutor by the Center for Academic Excellence on the basis of coursework performance across Intro to Python, Artificial Intelligence, Web Application Development, and Computer Networks & Security.
Undergraduate Research Assistant · Drew Summer Science Institute (DSSI) May 2026 – July 2026 · Advised by Prof. Alexander Rudniy, Drew University
The problem: general-purpose LLMs hallucinate on specialised particle-physics questions because Fermilab's technical knowledge is thinly represented in their training data. I built a complete pipeline to fix that — data collection, model adaptation, retrieval, and evaluation.
1️⃣ Corpus construction
- Built a BFS web crawler in Python (
Requests+BeautifulSoup) that harvested 936 of 1,000 public Fermilab pages — a 93.6% success rate, ~1.1 GB raw. - Cleaned and chunked the corpus into 8,847 passages, using document-level train/test splitting so no chunk from a test document could leak into training.
2️⃣ Synthetic supervision
- Generated 438 grounded question–answer pairs using Gemini 2.5 Flash as a teacher model.
- Applied 8-gram decontamination and quality filtering so the evaluation set stayed honest.
3️⃣ Model adaptation
- Fine-tuned Qwen3-4B-Instruct with QLoRA (4-bit NF4 quantization, LoRA rank 32) via Hugging Face TRL.
- Trained on the Purdue Anvil HPC cluster with SLURM job scheduling.
- Result: training loss down ~58%, token-level accuracy 56% → ~75%, with no overfitting signal in the validation curve.
4️⃣ Retrieval & delivery
- Shipped a Streamlit RAG application with hybrid retrieval — sparse BM25 and dense embeddings fused through Reciprocal Rank Fusion (RRF) over the 8,847-chunk index.
- Added a base-vs-fine-tuned comparison mode so reviewers can see the delta side by side.
- Built a Recall@k / MRR evaluation harness to measure retrieval quality quantitatively rather than by vibes.
📊 Results at a glance
| Metric | Value | Why it matters |
|---|---|---|
| Crawl success rate | 936 / 1,000 pages (93.6%) | Robust BFS crawler with retry and error handling, not a toy script |
| Raw corpus size | ~1.1 GB | Real-scale data engineering, not a sample notebook |
| Indexed passages | 8,847 chunks | Chunking strategy tuned for retrieval, not arbitrary splits |
| Synthetic QA pairs | 438 grounded pairs | Teacher-model supervision with provenance back to source text |
| Contamination control | 8-gram decontamination | Evaluation results are trustworthy, not inflated by leakage |
| Training loss reduction | ~58% | Fine-tuning genuinely converged |
| Token-level accuracy | 56% → ~75% | +19 points, with no overfitting in validation |
| Quantization | 4-bit NF4, LoRA rank 32 | A 4B model fine-tuned within realistic VRAM limits |
| Retrieval evaluation | Recall@k, MRR | Retrieval quality is measured, not assumed |
🧭 Research competencies demonstrated
| Area | Evidence from this project |
|---|---|
| Data engineering | BFS crawler, HTML cleaning, chunking, document-level train/test splitting |
| Experimental rigor | Leakage prevention, 8-gram decontamination, held-out evaluation set |
| Model adaptation | QLoRA / PEFT, 4-bit quantization, LoRA rank selection, loss-curve monitoring |
| Information retrieval | BM25 sparse + dense embeddings, Reciprocal Rank Fusion, Recall@k / MRR harness |
| HPC & tooling | SLURM job scripts, batch scheduling, multi-hour training runs on Purdue Anvil |
| Communication | Base-vs-fine-tuned comparison UI; presenting at Day of Scholars, Spring 2027 |
Python PyTorch Hugging Face Transformers TRL PEFT bitsandbytes sentence-transformers BM25 Streamlit Requests BeautifulSoup SLURM Gemini 2.5 Flash
Drew University — Madison, NJ
- B.S. Computer Science, Minor in Data Science — GPA: 3.75 / 4.0
- Expected Graduation: May 2027
- Merit Scholarship $15,000 (Fall 2025 – Spring 2026)
- Previously completed two years at Ho Chi Minh City University of Technology and Education, Vietnam
- Key Courses: Artificial Intelligence · Data Science · Software Engineering · Database Systems and Management · Statistics & Probability · Web Application Development · Computer Networks and Security
Drew University, Center for Academic Excellence · February 2026 – Present · Madison, NJ
- Selected as a Peer Subject Tutor on the strength of academic performance across CS coursework — Intro to Python, Artificial Intelligence, Web Application Development, and Computer Networks & Security.
- Deliver one-on-one and small-group tutoring in programming, algorithms, and data science concepts.
- Guide students through debugging, problem decomposition, and computational thinking rather than handing over answers.
- Also serve as a Foreign Language Tutor and Peer Academic Coach within the same center.
Ordered by technical depth. Each entry answers three questions: what it does, how it was built, and why it mattered.
Nov 2025 – Present
An interactive portfolio where visitors ask questions in natural language instead of scrolling.
- Built and deployed a responsive multi-page portfolio website using HTML, CSS, and JavaScript.
- Integrated an AI-powered interactive Chatbox so visitors can explore my background and projects conversationally.
- Implemented MCP (Model Context Protocol) for structured context passing and controlled, grounded AI responses.
- Applied the NLWeb protocol to expose site content as machine-readable knowledge for dynamic retrieval and reasoning.
- Managed project data through structured JSON schemas, making updates and scaling trivial.
🔗 Live: https://nguyentin2026.github.io/Project-Portfolio-Chatbox/ 📁 Repo: https://github.com/NguyenTin2026/Project-Portfolio-Chatbox
Aug 2025 – Oct 2025
A multi-source computer vision pipeline that runs live on consumer hardware.
- Built a real-time face detection pipeline with Python, OpenCV, and Haar Cascade classifiers, reaching ~95% detection accuracy on a custom labeled dataset.
- Engineered a live webcam inference system with on-frame annotation, holding 30 FPS on CPU and 60 FPS on GPU.
- Optimized frame processing and inference, cutting latency ~25% through preprocessing efficiency and performance tuning.
- Implemented multi-source input (webcam / image / video) behind a single interface.
📁 Repo: https://github.com/NguyenTin2026/Mini-Project-Computer-Vision-in-Deep-Learning
Aug 2025 – Present
- Real-time multi-class object detection using YOLOv5 + OpenCV.
- Data cleaning, exploratory analysis, and result visualization with Pandas, Seaborn, Matplotlib.
- Delivered as a client-facing freelance engagement.
📁 Repo: https://github.com/NguyenTin2026/Mini-Project-Computer-Vision-in-Deep-Learning
Oct 2025 – Present
- Built a full Streamlit web application for generating, previewing, and exporting a Curriculum Vitae.
- Developed in Python with a modular library structure for layout, data, and export logic.
- Deployed on Streamlit Cloud and publicly accessible — no login, no setup.
🔗 Live App: https://app-tutorial-buiding-cv-tintin.streamlit.app/
Oct 2025 – Present (originally explored Jun 2023 – Feb 2024)
- Applied data cleaning and preprocessing to handle missing values and outliers.
- Engineered meaningful features (family size, title extraction from names) that measurably lifted model performance.
- Built and evaluated Logistic Regression and Random Forest classifiers.
- Achieved ~85% accuracy on held-out test data.
- Queried and joined source data with PostgreSQL; visualized findings with Pandas, Matplotlib, Seaborn.
📁 Repo: https://github.com/NguyenTin2026/Data-Science-Tutorial-Step-By-Step-
Aug 2024 – Sep 2024
- Built REST APIs with FastAPI, including request validation and structured error responses.
- Tested endpoints systematically with Postman collections.
- Managed source control across Git, GitHub, and Bitbucket.
📁 Bitbucket: https://bitbucket.org/tintin_2026/workspace/overview
| Project | What it demonstrates | Stack | Link |
|---|---|---|---|
| Fermilab RAG + QLoRA Assistant | Full LLM lifecycle: crawl → chunk → synthesize → fine-tune → retrieve → evaluate | Qwen3-4B, TRL, PEFT, BM25+dense, SLURM | DSSI research — repo on request |
| Portfolio Website with AI Chatbox | Applied MCP + NLWeb for grounded, structured AI responses on a live site | HTML/CSS/JS, MCP, NLWeb, JSON | Live |
| Real-Time Face Detection | Latency-aware CV engineering on CPU and GPU | OpenCV, Haar Cascade | Repo |
| Object Detection (YOLOv5) | Modern detector deployment for a real client | YOLOv5, OpenCV | Repo |
| Streamlit CV Builder | Shipping and maintaining a public production app | Streamlit, Python | Live |
| Titanic Survival Prediction | Feature engineering and honest model comparison | scikit-learn, Pandas | Repo |
| API Testing & Backend | Backend fundamentals and API contract testing | FastAPI, Postman | Bitbucket |
| Credential | Issuer | Date | What it covers |
|---|---|---|---|
| Mastering Claude AI: Prompting, APIs, RAG, and MCP | Edureka (via Coursera) | Jul 2026 | Prompt engineering, LLM API integration, retrieval-augmented generation, Model Context Protocol — directly applied in my portfolio chatbox and Fermilab RAG work |
| Engineer AI Agents with Agent Development Kit (ADK) Skill Badge — Intermediate |
Google Cloud | May 2026 | Building, orchestrating, and deploying multi-step AI agents on Google Cloud |
| CRLA International Tutor Training Program (ITTPC), Level I | Drew University (CRLA-accredited) | Spring 2026 | Internationally recognised peer-tutoring certification: learning theory, session structuring, active listening, ethics |
🎯 Currently working toward: deeper MLOps tooling (MLflow, Weights & Biases) and LLM evaluation methodology beyond Recall@k / MRR.
Text version first — recruiters and ATS parsers read this; badges below are for humans.
| Category | Skills |
|---|---|
| Programming Languages | Python · R · SQL (PostgreSQL) · Java · C++ · JavaScript · TypeScript · HTML · CSS |
| LLM & Generative AI | LLM fine-tuning (QLoRA / LoRA via PEFT) · 4-bit quantization (bitsandbytes) · Hugging Face Transformers & TRL · Retrieval-Augmented Generation (RAG) · Hybrid retrieval (BM25 + dense) · Reciprocal Rank Fusion · sentence-transformers · Prompt engineering · Model Context Protocol (MCP) · NLWeb |
| Machine Learning | PyTorch · TensorFlow · Keras · Scikit-learn · Feature engineering · Model evaluation (Recall@k, MRR, accuracy, loss curves) · Logistic Regression · Random Forest |
| Computer Vision | OpenCV · Haar Cascade classifiers · YOLOv5 · Real-time inference optimization · Multi-source input pipelines |
| Data Science & Analysis | NumPy · Pandas · SciPy · Matplotlib · Seaborn · Excel · Data cleaning · Outlier & missing-value handling · Exploratory data analysis |
| Data Engineering | Web crawling (Requests, BeautifulSoup, BFS traversal) · Corpus chunking · Document-level train/test splitting · Data decontamination (n-gram) |
| Web & Backend | Streamlit · FastAPI · REST API design · Postman testing · Responsive front-end (HTML/CSS/JS) · JSON schema design |
| Cloud & Infrastructure | Google Cloud Platform (GCP) · Microsoft Azure · Docker · SLURM (HPC job scheduling) · Streamlit Cloud · GitHub Pages |
| Databases | PostgreSQL · SQL querying & joins |
| Cybersecurity | Kali Linux · Burp Suite · TryHackMe practical labs |
| Dev Environments | Jupyter Notebook · Google Colab · VS Code · PyCharm · Anaconda · Sublime Text · Linux |
| Version Control | Git · GitHub · GitLab · Bitbucket |
| Soft Skills | Peer tutoring (CRLA Level I certified) · Technical explanation to non-experts · Research communication · Independent project ownership · Client-facing freelance delivery |
-2C3E50?style=for-the-badge&logoColor=white){kind=icon}
May 2023 – May 2025
- Founded 2 free English-learning platforms on Facebook, growing to 900+ followers.
- Delivered weekly tutoring in English and Math and mentored learners one-on-one at no cost.
Jan 2021 – May 2025
- Managed TikTok (22K followers, 2M+ views), Instagram, and Facebook channels.
- Produced multimedia content and lifted average view duration by 25% through iterative format testing — the same measure-then-optimize instinct I apply to models.
Now
- 🔭 LangChain / LlamaIndex — agentic and multi-step retrieval orchestration
- 🧪 Advanced MLOps — experiment tracking, model registries, reproducible training
- 📏 LLM evaluation — going beyond Recall@k/MRR toward faithfulness and groundedness metrics
Next
- 🌱 Generative AI at production scale — serving, caching, and cost control
- ⚙️ Distributed training and inference optimization on HPC
- 👯 Open to collaboration on AI automation, production ML, and open-source projects
- 💡 Planning technical write-ups on Medium / Dev.to — starting with the Fermilab RAG pipeline
- I finish things. Every project listed above has a repo, a live URL, or measured results — not a half-finished notebook.
- I measure what I build. Recall@k, MRR, token-level accuracy, FPS, latency — I report numbers, not adjectives.
- I work across the whole stack. Crawler → dataset → fine-tune → retrieval → deployed UI, on both laptops and HPC clusters.
- I teach. As a CRLA-certified peer tutor, I explain complex ideas clearly — useful on any team.
- I'm available. Actively seeking Data Science / ML Engineering internships for 2026–2027.
💻 Developer & Code Hosting Platforms
🤖 AI / ML Platforms I Work With
☁️ Cloud & Deployment Platforms
🌐 Social Media
📧 Email: ndo1@drew.edu 📱 Phone: available on request via email or LinkedIn
Thanks for stopping by! Browse the repos, open an issue, or reach out — I'm always up for a good AI/ML collaboration. 😊